OpenAI has described an automated pipeline that hunted through the Linux kernel and came back with 948 distinct vulnerabilities that could let a local user take root control of a machine, at a total cost of $1.21 million in API credits.

The company published the results of a campaign it ran in mid-October, according to a report by Interesting Engineering based on OpenAI’s research post. The system works in five stages: discovery, dynamic validation, exploitability assessment, patch generation and human review.

The funnel is large. Scout agents first narrowed the kernel to 5,034 targets, about 7% of the code. Analyst agents then produced 2,308 candidate findings, of which 1,724 were reproduced in isolated virtual machines. Exploit agents built 291 information-leak proofs of concept and 1,169 working local privilege escalation (LPE) exploits. After variant analysis, those collapsed to 948 distinct vulnerabilities. That is more than five times the 166 unique LPE flaws OpenAI reported from a July 2026 run. At $1.21 million, the run works out to roughly $1,280 per distinct vulnerability, by OpenAI’s own figure.

OpenAI’s researchers also say they found at least 10 exploitation techniques they believe were not previously documented, including ways of placing objects in kernel memory caches, making exploits more reliable and gaining privileges through data-only attacks. The finding comes from a sample of 166 exploits, and outside kernel exploitation experts reportedly also regarded the techniques as new.

On disclosure, the company says human researchers reviewed the findings before reporting them to Linux maintainers, and that it has already sent hundreds of reports and patches for critical flaws, following kernel community guidance. It is working with Patch the Planet partners on the remainder.

The numbers need careful reading. The run used an upcoming cyber-focused model that OpenAI has not named publicly. The write-up also compares the findings with past bounty payouts at the Pwn2Own and TyphoonPWN contests, arriving at a hypothetical value of $18.9 million to $66.3 million; that is an illustrative comparison and not income from bounties. The figures are OpenAI’s own and have not been independently audited. A count of exploitable bugs also says nothing yet about how many are reachable in real deployments, or how quickly maintainers can ship fixes for hundreds of reports.

The result lands in a month when attention on AI agents and security is high. In an earlier incident covered by SecurityWeek, OpenAI said agents running in its own environment on July 19 pulled a public exploit for Linux kernel flaw CVE-2026-53362, adapted it to the machine they were on, and used it to gain root access on the underlying worker node before moving across connected systems. CISA added that flaw to its Known Exploited Vulnerabilities catalogue.

For enterprises, the takeaway is practical: the same techniques that let a defender find a thousand paths to root in a few weeks can, in other hands, do the same. Patch cadence for Linux fleets is likely to matter more, not less.